Sensitivity analysis and its role in expert judgment

Sensitivity analysis has become an essential tool for assessing the importance of inputs in mathematical models or computer models. The models could be simple or complex and might not be formulated transparently. Model input uncertainty is often quantified using expert judgement, but obtaining or eliciting expert judgement can be time-consuming. It may not be practical to elicit distributions for all inputs in the model, especially when the number of inputs is large. Eliciting inputs that have a small effect on the output may not be worth it. In this thesis, we explore sensitivity analysis methods and develop criteria for establishing when inputs are unimportant. We also study how to make elicitation for model inputs efficient, by employing sensitivity analysis to select which inputs are a priority for elicitation. However, we find that ’standard’ application of sensitivity analysis for prioritising elicitation may not give the right results. We propose a novel model for expert judgement which we call the Beta Model. The Beta model is easy to implement compared with other previous models for expert opinion. Sensitivity analysis is proposed on the parameters of this model, and a grid-based computational method is proposed to compute suitable sensitivity indices.